UniStore: A fault-tolerant marriage of causal and strong consistency
summary
UniStore is a transactional POR store that actually implements optimistic certification using two‑phase commit + Paxos and a lightweight uniformity check that forces a strong transaction to depend only on causal updates replicated to f+1 data centers. A prototype in Erlang runs on Amazon EC2 and is benchmarked with RUBiS; with a 15% update workload (≈10% strong tx) UniStore achieves throughput at saturation 72% and 183% higher than red‑blue and all‑strong baselines, averages 16.5 ms latency vs 80.4 ms for all‑strong, while causal transactions stay ~1.2 ms.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:21Geo-replication
- 0:42Fundamental trade-offs
- 1:12The hybrid alternative
- 2:01A banking application
- 6:16UniStore: causal baseline
- 6:35UniStore: strong txs
- 8:29US data UniStore: key challenge center crashes replicating to
- 9:25Related work
- 9:55UniStore: uniformity
- 10:50UniStore: other features
- 11:46Evaluation: results
- 14:01Conclusion
related talks
41mTransactions: myths, surprises and opportunities
46mKEEPING UP IN A DATA-INTENSIVE WORLD WITH SINGLESTORE
14mBolt-On Strong Consistency: Specification, Implementation, and Verification
15mRethinking Safe Consistency in Distributed Object-Oriented Programming
19mBuilding a Performant Configuration-Store Layer
22mHow to Enforce and Verify Invariants in Weakly Consistent Databases
01h 00mBeyond Eventual Consistency
46mBeyond Eventual Consistency